Fig. 1: The workflow of KIDA, each subgraph is a step.
From: Knowledge-based inductive bias and domain adaptation for cell type annotation
a Unified feature space as genes. b Retrieve pathway names based on gene names. c Mask the parameters of the linear transformation using gene-pathway relationships. d Transform features from genes to functional patterns. e First round of supervised learning: learn the interaction of functional patterns on the reference dataset and add a token as cell embedding. f Second round of supervised learning: on the query dataset, the model from step (e) selects anchors (cells with high confidence) to form a new training set. Then, the model utilizes pseudo labels of the anchors for self-knowledge distillation. Orange arrows are training, gray arrows are inference.
